Sorting & Bin Picking

Application

What is bin picking?

Bin picking — the targeted picking of individual workpieces from an unordered bin (small load carrier, mesh box, cardboard box) — is the most technically demanding handling task in industrial assembly. Workpieces lie chaotically, overlap, jam, are partially hidden. The robot must recognise: which part can I pick next without damaging neighbours and without touching the box?

Before 2020, this was practically unsolvable with classical 2D vision. Only with the broad breakthrough of deep learning for 3D point clouds and the availability of robust industrial 3D sensors did bin picking become economically usable. Today (2026) it is a productive standard application — for electronics feeding, packaging, order picking, recycling and many more.

The economic benefit: bin picking makes classical feeding systems obsolete. Instead of vibratory bowls, oscillating feeders or elaborately fabricated trays, the robot can pick directly from the delivered small load carrier. This saves investment, floor space and makes part changeovers orders of magnitude faster.

Principle of the bin-picking cell

1. 3D capture — the camera captures the current scene as a point cloud. Depending on the system 1-5 seconds capture time.
2. Object recognition — the software identifies individual workpieces (instance segmentation) and their pose in space (6DoF: X/Y/Z + rotation).
3. Grasp selection — the algorithm calculates the best grasp point: which part lies most favourably, where does the gripper fit, are there collisions with other parts or bin walls?
4. Path planning — the robot plans its motion so that gripper and workpiece do not touch the container.
5. Pick and place — robot picks, moves the part to the placing station or directly into downstream processing.

In modern systems this cycle runs in under 5 seconds. For most applications 4-8 seconds per part are enough, critical cells reach under 3 seconds.

3D sensor technologies

• Structured light (Photoneo PhoXi, Zivid) — projects a pattern, measures deformation. Very precise, robust in the near range (up to 2 m), often blue light for contrast on metal parts.
• Stereo vision (Roboception rc_visard, Basler blaze) — two cameras compute depth. Passive (no projector), cheaper, but more sensitive on uniform surfaces.
• Laser triangulation (SICK Ruler, LMI Gocator) — very precise (0.01 mm), but slower scan. For high-precision small parts.
• Time-of-flight (ToF) — fast, large range, but lower resolution. Rather for coarse palletizing tasks.

Sorting applications (not just from bins)

The same principle can be applied to sorting tasks — objects are not picked from a bin but sorted from a conveyor by characteristics:

• Recycling and material separation — by material, colour, size. AI vision also recognises heavily damaged parts.
• Logistics and order picking — sort parcels by destination ramp. Label reading and volume measurement in parallel.
• Agriculture and food — sort fruit/vegetables by ripeness, size, quality.
• Return/reverse logistics — automatically classify and re-sort e-commerce returns.

Gripping concepts

The gripper often decides success more than the vision. Standard options:

• Vacuum cups (single or multi-cup) — for flat surfaces, boxes, blisters. Fast, cheap, but sensitive to oil and contamination.
• Parallel gripper (2-finger) — for standard workpieces with clear grip geometries. Robust.
• 3-finger adaptive (Robotiq 3F, Schunk EGP-C) — conforms to the workpiece geometry, for irregular shapes.
• Magnetic gripper — for ferromagnetic parts, insensitive to oil.
• Soft gripper (silicone fingers, fin-ray) — for delicate or irregular parts (food, plastic moulded parts).

For flexible applications, tool changers are increasingly used — the robot swaps the matching gripper depending on the part.

Typical applications

• Electronics assembly — parts from trays or bins directly to the assembly line (connectors, screws, housing parts).
• Consumer goods packaging — products from bins into blister or carton packaging.
• Automotive feeding — small-parts supply to assembly lines without elaborate vibratory bowls.
• Life science & pharma — pick vials from load carriers (with vacuum cups, careful with fragility).
• Recycling — automated material sorting on conveyors (PET, PP, metals).
• Logistics / order picking — pick goods for orders from shelves (order picking, classically the hardest case due to extreme part variety).

Common challenges

• Reflective and transparent parts — hard for 3D sensors to capture. Matting spray, polarising filters or special sensors required. For glass and high-gloss metal, currently still difficult.
• Ambient light — with structured light, stray light must be controlled. Cell safeguarding is also light safeguarding.
• Jammed or overlapping parts — the last remainder in the bin is often the hardest. Vibrators or tilt stations can help, but 100 % pick rate is rarely economical.
• Cycle time — for high-cycle applications (under 3 seconds per pick), bin picking is often not enough — classical feeding is faster there. Bin picking plays out its strength at medium cycle rates and high variant variety.
• Part variety and changeover — CAD matching needs CAD data. Deep-learning systems need training images — depending on the vendor between 10 and 500 examples. Changeover takes hours instead of days (compared to classical feeding), but is not entirely zero.
• Scrap and mispick rate — 0.5-2 % mispicks are standard. The cell needs a handler for mispicks (reject chute, retry, alarm).